The power of AI — on your visitors’ machine, at no extra cost.
Embed an assistant that runs entirely in the browser, on the visitor’s own GPU. No server, no per-token bill, no data leaving the device. You shape what it does with a single prompt.
Zero inference cost
The compute runs on your visitor’s GPU, not a server. No per-token bill, no rate limits, infinite scale. Flat price, unlimited usage.
Private by design
The data never leaves the browser. Nothing transits through us — a decisive argument for privacy-sensitive sites (health, legal, HR).
No server, no ops
No GPU fleet to run, no scaling to manage, no API keys to rotate. The model streams once from a CDN and is cached on the device.
Shaped by a prompt
You define what the AI does with a predefined instruction — its role, its logic, its guardrails. No fine-tuning, no data scientist required.
Try it — classify, extract & chat, on your GPU
An ultra-light model (LFM2.5 230M) classifies, extracts and chats — in French too — shaped by a prompt, running entirely in your browser (no server). First run downloads it once (~149 MB), then it’s cached & offline.
How it works
Drop one <script> tag and write the instruction that shapes the assistant.
On first use, a lightweight model streams from a CDN and caches on the device (then it’s instant, even offline).
Every prompt runs on the visitor’s GPU through our WebGPU kernels — nothing is sent anywhere.
What it’s great at
A light on-device model shines on scoped tasks — the ones a good prompt fully defines:
Embed it on your site — SDK v0
The SDK is live. One script tag mounts the assistant; a prompt shapes it. The model only downloads when a visitor opens the widget — your page score is untouched.
<script src="https://brimkern.com/sdk.js"></script>
<script>
Brimkern.embed({
system: 'You are a friendly, concise support assistant for Acme.',
});
</script>See a live integration v0 — chat widget, LFM2 .brik model URL, colors & wording, few-shot examples, and knowledge documents (answers from YOUR content, selected locally, nothing sent anywhere). Tools are next. On the default 230M model, keep notes short and factual: it quotes them well, but it can mix up two numbers sitting in the same paragraph.
Our own engine, not a black box
Hand-written WebGPU kernels run the model; weights ship in our compact BRIK format, quantized down to int3 and streamed in seconds, then cached. A model is picked to fit — the loader even reads the visitor’s GPU and connection to recommend the right one.
Runs in modern browsers with WebGPU (Chrome, Edge, Safari, Firefox). Best for well-scoped tasks — not open-ended reasoning. First load downloads the model once; every visit after is instant and offline-capable.
Brimkern — local WebGPU inference. The embeddable SDK (v0) is live and free — the engine is open source (MIT).
Brimkern — open WebGPU engine, built by Romain Khanoyan. Local AI, WebGPU, on-device engines.